Denoising Pretraining

Recent work demonstrates the potential of training one model for multilingual machine translation. In parallel, denoising pretraining using unlabeled monolingual data as a starting point for finetuning bitext machine translation …...